API: Improved inference of datetime/timedelta with mixed null objects. (GH7431)

[jreback]

• regression from 0.13.1 in interpretation of an object Index
• additional tests to validate datetimelike inferences

closes #7431

[jreback]

@cpcloud @jorisvandenbossche have a look

this inference biz is pretty complicated; hopefully sightly less complicated now (issue is that you can easily mess up perf if you keep trying to infer object dtypes).

technically I did change one API here, in that

In 0.14.0

In [1]: from datetime import timedelta

In [2]: Series([timedelta(1)])
Out[2]: 
0   1 days
dtype: timedelta64[ns]

In [3]: DataFrame([timedelta(1)])
Out[3]: 
                0
0  1 day, 0:00:00

In [4]: DataFrame([timedelta(1)]).dtypes
Out[4]: 
0    object
dtype: object

Now

In [2]: DataFrame([timedelta(1)])
Out[2]: 
       0
0 1 days

In [3]: DataFrame([timedelta(1)]).dtypes
Out[3]: 
0    timedelta64[ns]
dtype: object

So was not tested before

[jorisvandenbossche]

I can't comment on the code, but for the API change, it was indeed no problem to merge I think as it was actually more a bug that is fixed. And not one that too many people will rely upon I think, as eg with DataFrame({'A':[timedelta(1), timedelta(2)]}) it already did work as expected before.